| <!DOCTYPE html> |
| |
| <!--- |
| Licensed to the Apache Software Foundation (ASF) under one |
| or more contributor license agreements. See the NOTICE file |
| distributed with this work for additional information |
| regarding copyright ownership. The ASF licenses this file |
| to you under the Apache License, Version 2.0 (the |
| "License"); you may not use this file except in compliance |
| with the License. You may obtain a copy of the License at |
| http://www.apache.org/licenses/LICENSE-2.0 |
| Unless required by applicable law or agreed to in writing, |
| software distributed under the License is distributed on an |
| "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
| KIND, either express or implied. See the License for the |
| specific language governing permissions and limitations |
| under the License. |
| --> |
| |
| <html lang=" en"><head> |
| <meta charset="utf-8"> |
| <meta http-equiv="X-UA-Compatible" content="IE=edge"> |
| <meta name="viewport" content="width=device-width, initial-scale=1"> |
| <link href="/versions/1.9.1/assets/img/mxnet-icon.png" rel="icon" type="image/png"><!-- Begin Jekyll SEO tag v2.6.1 --> |
| <title>Data Parallelism with Multiple CPU/GPUs on MXNet | Apache MXNet</title> |
| <meta name="generator" content="Jekyll v3.8.6" /> |
| <meta property="og:title" content="Data Parallelism with Multiple CPU/GPUs on MXNet" /> |
| <meta property="og:locale" content="en_US" /> |
| <meta name="description" content="A flexible and efficient library for deep learning." /> |
| <meta property="og:description" content="A flexible and efficient library for deep learning." /> |
| <link rel="canonical" href="https://mxnet.apache.org/versions/1.9.1/api/faq/multi_device" /> |
| <meta property="og:url" content="https://mxnet.apache.org/versions/1.9.1/api/faq/multi_device" /> |
| <meta property="og:site_name" content="Apache MXNet" /> |
| <script type="application/ld+json"> |
| {"description":"A flexible and efficient library for deep learning.","headline":"Data Parallelism with Multiple CPU/GPUs on MXNet","@type":"WebPage","url":"https://mxnet.apache.org/versions/1.9.1/api/faq/multi_device","@context":"https://schema.org"}</script> |
| <!-- End Jekyll SEO tag --> |
| <link rel="stylesheet" href="/versions/1.9.1/assets/docsearch.min.css" /><link rel="stylesheet" href="/versions/1.9.1/assets/main.css"><link type="application/atom+xml" rel="alternate" href="https://mxnet.apache.org/versions/1.9.1/feed.xml" title="Apache MXNet" /><!-- Matomo --> |
| <script> |
| var _paq = window._paq = window._paq || []; |
| /* tracker methods like "setCustomDimension" should be called before "trackPageView" */ |
| /* We explicitly disable cookie tracking to avoid privacy issues */ |
| _paq.push(['disableCookies']); |
| _paq.push(['trackPageView']); |
| _paq.push(['enableLinkTracking']); |
| (function() { |
| var u="https://analytics.apache.org/"; |
| _paq.push(['setTrackerUrl', u+'matomo.php']); |
| _paq.push(['setSiteId', '23']); |
| var d=document, g=d.createElement('script'), s=d.getElementsByTagName('script')[0]; |
| g.async=true; g.src=u+'matomo.js'; s.parentNode.insertBefore(g,s); |
| })(); |
| </script> |
| <!-- End Matomo Code --> |
| |
| <script src="/versions/1.9.1/assets/js/jquery-3.3.1.min.js"></script> |
| <script src="/versions/1.9.1/assets/js/docsearch.min.js"></script><script src="/versions/1.9.1/assets/js/globalSearch.js" defer></script> |
| <script src="/versions/1.9.1/assets/js/clipboard.js" defer></script> |
| <script src="/versions/1.9.1/assets/js/copycode.js" defer></script></head> |
| <body><header class="site-header" role="banner"> |
| |
| <script> |
| $(document).ready(function () { |
| |
| // HEADER OPACITY LOGIC |
| |
| function opacity_header() { |
| var value = "rgba(4,140,204," + ($(window).scrollTop() / 300 + 0.4) + ")" |
| $('.site-header').css("background-color", value) |
| } |
| |
| $(window).scroll(function () { |
| opacity_header() |
| }) |
| opacity_header(); |
| |
| // MENU SELECTOR LOGIC |
| $('.page-link').each( function () { |
| if (window.location.href.includes(this.href)) { |
| $(this).addClass("page-current"); |
| } |
| }); |
| }) |
| </script> |
| <div class="wrapper"> |
| <a class="site-title" rel="author" href="/versions/1.9.1/"><img |
| src="/versions/1.9.1/assets/img/mxnet_logo.png" class="site-header-logo"></a> |
| <nav class="site-nav"> |
| <input type="checkbox" id="nav-trigger" class="nav-trigger"/> |
| <label for="nav-trigger"> |
| <span class="menu-icon"> |
| <svg viewBox="0 0 18 15" width="18px" height="15px"> |
| <path d="M18,1.484c0,0.82-0.665,1.484-1.484,1.484H1.484C0.665,2.969,0,2.304,0,1.484l0,0C0,0.665,0.665,0,1.484,0 h15.032C17.335,0,18,0.665,18,1.484L18,1.484z M18,7.516C18,8.335,17.335,9,16.516,9H1.484C0.665,9,0,8.335,0,7.516l0,0 c0-0.82,0.665-1.484,1.484-1.484h15.032C17.335,6.031,18,6.696,18,7.516L18,7.516z M18,13.516C18,14.335,17.335,15,16.516,15H1.484 C0.665,15,0,14.335,0,13.516l0,0c0-0.82,0.665-1.483,1.484-1.483h15.032C17.335,12.031,18,12.695,18,13.516L18,13.516z"/> |
| </svg> |
| </span> |
| </label> |
| <div class="gs-search-border"> |
| <div id="gs-search-icon"></div> |
| <form id="global-search-form"> |
| <input id="global-search" type="text" title="Search" placeholder="Search" /> |
| <div id="global-search-dropdown-container"> |
| <button class="gs-current-version btn" type="button" data-toggle="dropdown"> |
| <span id="gs-current-version-label">1.9.1</span> |
| <svg class="gs-dropdown-caret" viewBox="0 0 32 32" class="icon icon-caret-bottom" aria-hidden="true"> |
| <path class="dropdown-caret-path" d="M24 11.305l-7.997 11.39L8 11.305z"></path> |
| </svg> |
| </button> |
| <ul class="gs-opt-group gs-version-dropdown"> |
| |
| |
| <li class="gs-opt gs-versions">master</li> |
| |
| |
| |
| <li class="gs-opt gs-versions active">1.9.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.8.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.7.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.6.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.5.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.4.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.3.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.2.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.1.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.0.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">0.12.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">0.11.0</li> |
| |
| |
| </ul> |
| </div> |
| <span id="global-search-close">x</span> |
| </form> |
| </div> |
| <div class="trigger"> |
| <div id="global-search-mobile-border"> |
| <div id="gs-search-icon-mobile"></div> |
| <input id="global-search-mobile" placeholder="Search..." type="text"/> |
| <div id="global-search-dropdown-container-mobile"> |
| <button class="gs-current-version-mobile btn" type="button" data-toggle="dropdown"> |
| <svg class="gs-dropdown-caret" viewBox="0 0 32 32" class="icon icon-caret-bottom" aria-hidden="true"> |
| <path class="dropdown-caret-path" d="M24 11.305l-7.997 11.39L8 11.305z"></path> |
| </svg> |
| </button> |
| <ul class="gs-opt-group gs-version-dropdown-mobile"> |
| |
| |
| <li class="gs-opt gs-versions">master</li> |
| |
| |
| |
| <li class="gs-opt gs-versions active">1.9.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.8.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.7.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.6.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.5.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.4.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.3.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.2.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.1.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">1.0.0</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">0.12.1</li> |
| |
| |
| |
| <li class="gs-opt gs-versions">0.11.0</li> |
| |
| |
| </ul> |
| </div> |
| </div> |
| <a class="page-link" href="/versions/1.9.1/get_started">Get Started</a> |
| <a class="page-link" href="/versions/1.9.1/features">Features</a> |
| <a class="page-link" href="/versions/1.9.1/ecosystem">Ecosystem</a> |
| <a class="page-link" href="/versions/1.9.1/api">Docs & Tutorials</a> |
| <a class="page-link" href="/versions/1.9.1/trusted_by">Trusted By</a> |
| <a class="page-link" href="https://github.com/apache/mxnet">GitHub</a> |
| <div class="dropdown" style="min-width:100px"> |
| <span class="dropdown-header">Apache |
| <svg class="dropdown-caret" viewBox="0 0 32 32" class="icon icon-caret-bottom" aria-hidden="true"><path class="dropdown-caret-path" d="M24 11.305l-7.997 11.39L8 11.305z"></path></svg> |
| </span> |
| <div class="dropdown-content" style="min-width:250px"> |
| <a href="https://www.apache.org/foundation/">Apache Software Foundation</a> |
| <a href="https://www.apache.org/licenses/">License</a> |
| <a href="/versions/1.9.1/api/faq/security.html">Security</a> |
| <a href="https://privacy.apache.org/policies/privacy-policy-public.html">Privacy</a> |
| <a href="https://www.apache.org/events/current-event">Events</a> |
| <a href="https://www.apache.org/foundation/sponsorship.html">Sponsorship</a> |
| <a href="https://www.apache.org/foundation/thanks.html">Thanks</a> |
| </div> |
| </div> |
| <div class="dropdown"> |
| <span class="dropdown-header">1.9.1 |
| <svg class="dropdown-caret" viewBox="0 0 32 32" class="icon icon-caret-bottom" aria-hidden="true"><path class="dropdown-caret-path" d="M24 11.305l-7.997 11.39L8 11.305z"></path></svg> |
| </span> |
| <div class="dropdown-content"> |
| <a href="/">master</a> |
| <a class="dropdown-option-active" href="/versions/1.9.1/">1.9.1</a> |
| <a href="/versions/1.8.0/">1.8.0</a> |
| <a href="/versions/1.7.0/">1.7.0</a> |
| <a href="/versions/1.6.0/">1.6.0</a> |
| <a href="/versions/1.5.0/">1.5.0</a> |
| <a href="/versions/1.4.1/">1.4.1</a> |
| <a href="/versions/1.3.1/">1.3.1</a> |
| <a href="/versions/1.2.1/">1.2.1</a> |
| <a href="/versions/1.1.0/">1.1.0</a> |
| <a href="/versions/1.0.0/">1.0.0</a> |
| <a href="/versions/0.12.1/">0.12.1</a> |
| <a href="/versions/0.11.0/">0.11.0</a> |
| </div> |
| </div> |
| </div> |
| </nav> |
| </div> |
| </header> |
| <main class="page-content" aria-label="Content"> |
| <script> |
| |
| </script> |
| <article class="post"> |
| |
| <header class="post-header wrapper"> |
| <h1 class="post-title">Data Parallelism with Multiple CPU/GPUs on MXNet</h1> |
| <h3></h3></header> |
| |
| <div class="post-content"> |
| <div class="wrapper"> |
| <div class="row"> |
| <div class="col-3 docs-side-bar"> |
| <h3 style="text-transform: capitalize; padding-left:10px">faq</h3> |
| <ul> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/add_op_in_backend">A Beginner's Guide to Implementing Operators in MXNet Backend</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/caffe">Convert from Caffe to MXNet</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/cloud">MXNet on the Cloud</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/distributed_training">Distributed Training in MXNet</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/env_var">Environment Variables</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/float16">Float16</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/large_tensor_support">Using MXNet with Large Tensor Support</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/model_parallel_lstm">Model Parallel</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/multi_device">Data Parallelism with Multiple CPU/GPUs on MXNet</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/new_op">Create New Operators</a></li> |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/nnpack">NNPACK for Multi-Core CPU Support in MXNet</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/perf">Some Tips for Improving MXNet Performance</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/recordio">Create a Dataset Using RecordIO</a></li> |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/s3_integration">Use data from S3 for training</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/security">MXNet Security Best Practices</a></li> |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/smart_device">Deep Learning at the Edge</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/visualize_graph">Visualize Neural Networks</a></li> |
| <!-- page-category --> |
| |
| |
| <li><a href="/versions/1.9.1/api/faq/why_mxnet">Why MXNet came to be?</a></li> |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| |
| <!-- page-category --> |
| <!-- resource-p --> |
| </ul> |
| </div> |
| <div class="col-9"> |
| <!--- Licensed to the Apache Software Foundation (ASF) under one --> |
| |
| <!--- or more contributor license agreements. See the NOTICE file --> |
| |
| <!--- distributed with this work for additional information --> |
| |
| <!--- regarding copyright ownership. The ASF licenses this file --> |
| |
| <!--- to you under the Apache License, Version 2.0 (the --> |
| |
| <!--- "License"); you may not use this file except in compliance --> |
| |
| <!--- with the License. You may obtain a copy of the License at --> |
| |
| <!--- http://www.apache.org/licenses/LICENSE-2.0 --> |
| |
| <!--- Unless required by applicable law or agreed to in writing, --> |
| |
| <!--- software distributed under the License is distributed on an --> |
| |
| <!--- "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY --> |
| |
| <!--- KIND, either express or implied. See the License for the --> |
| |
| <!--- specific language governing permissions and limitations --> |
| |
| <!--- under the License. --> |
| |
| <h1 id="run-mxnet-on-multiple-cpu-gpus-with-data-parallelism">Run MXNet on Multiple CPU/GPUs with Data Parallelism</h1> |
| |
| <p><em>MXNet</em> supports training with multiple CPUs and GPUs, which may be located on different physical machines.</p> |
| |
| <h2 id="data-parallelism-vs-model-parallelism">Data Parallelism vs Model Parallelism</h2> |
| |
| <p>By default, <em>MXNet</em> uses data parallelism to partition the workload over multiple |
| devices. |
| Assume there are <em>n</em> devices. |
| Then each one will receive a copy of the complete model |
| and train it on <em>1/n</em> of the data. |
| The results such as gradients and |
| updated model are communicated across these devices.</p> |
| |
| <p>MXNet also supports model parallelism. |
| In this approach, each device holds onto only part of the model. |
| This proves useful when the model is too large to fit onto a single device. |
| As an example, see the following <a href="./model_parallel_lstm.md">tutorial</a> |
| which shows how to use model parallelism for training a multi-layer LSTM model. |
| In this tutorial, we'll focus on data parallelism.</p> |
| |
| <h2 id="multiple-gpus-within-a-single-machine">Multiple GPUs within a Single Machine</h2> |
| |
| <h3 id="workload-partitioning">Workload Partitioning</h3> |
| |
| <p>By default, <em>MXNet</em> partitions a data batch evenly among the available GPUs. |
| Assume a batch size <em>b</em> and assume there are <em>k</em> GPUs, then in one iteration |
| each GPU will perform forward and backward on <em>b/k</em> examples. The |
| gradients are then summed over all GPUs before updating the model.</p> |
| |
| <h3 id="how-to-use">How to Use</h3> |
| |
| <blockquote> |
| <p>To use GPUs, we need to compile MXNet with GPU support. For |
| example, set <code>USE_CUDA=1</code> in <code>config.mk</code> before <code>make</code>. (see |
| <a href="/get_started">MXNet installation guide</a> for more options).</p> |
| </blockquote> |
| |
| <p>If a machine has one or more GPU cards installed, |
| then each card is labeled by a number starting from 0. |
| To use a particular GPU, one can either |
| specify the context <code>context</code> in code |
| or pass <code>--gpus</code> at the command line. |
| For example, to use GPU 0 and 2 in python, |
| one can typically create a module with |
| <code>python |
| import mxnet as mx |
| module = mx.module.Module(context=[mx.gpu(0), mx.gpu(2)], ...) |
| </code> |
| while if the program accepts a <code>--gpus</code> flag (as seen in |
| <a href="https://github.com/apache/mxnet/tree/v1.x/example/image-classification">example/image-classification</a>), |
| then we can try |
| <code>bash |
| python train_mnist.py --gpus 0,2 ... |
| </code></p> |
| |
| <h3 id="advanced-usage">Advanced Usage</h3> |
| |
| <p>If the available GPUs are not all equally powerful, |
| we can partition the workload accordingly. |
| For example, if GPU 0 is 3 times faster than GPU 2, |
| then we might use the workload option <code>work_load_list=[3, 1]</code>, |
| see <a href="/api/python/docs/api/module/index.html">Module</a> |
| for more details.</p> |
| |
| <p>Training with multiple GPUs should yield the same results |
| as training on a single GPU if all other hyper-parameters are the same. |
| In practice, the results may exhibit small differences, |
| owing to the randomness of I/O (random order or other augmentations), |
| weight initialization with different seeds, and CUDNN.</p> |
| |
| <p>We can control on which devices the gradient is aggregated |
| and on which device the model is updated via <a href="/api/python/docs/api/kvstore/index.html"><code>KVStore</code></a>, |
| the <em>MXNet</em> module that supports data communication. |
| One can either use <code>mx.kvstore.create(type)</code> to get an instance |
| or use the program flag <code>--kv-store type</code>.</p> |
| |
| <p>There are two commonly used types,</p> |
| |
| <ul> |
| <li><code>local</code>: all gradients are copied to CPU memory and weights are updated there.</li> |
| <li><code>device</code>: both gradient aggregation and weight updates are run on GPUs. |
| With this setting, the <code>KVStore</code> also attempts to use GPU peer-to-peer communication, |
| potentially accelerating the communication. |
| Note that this option may result in higher GPU memory usage.</li> |
| </ul> |
| |
| <p>When using a large number of GPUs, e.g. >=4, we suggest using <code>device</code> for better performance.</p> |
| |
| <h2 id="distributed-training-with-multiple-machines">Distributed Training with Multiple Machines</h2> |
| |
| <p><code>KVStore</code> also supports a number of options for running on multiple machines.</p> |
| |
| <ul> |
| <li><code>dist_sync</code> behaves similarly to <code>local</code> but exhibits one major difference. |
| With <code>dist_sync</code>, <code>batch-size</code> now means the batch size used on each machine. |
| So if there are <em>n</em> machines and we use batch size <em>b</em>, |
| then <code>dist_sync</code> behaves like <code>local</code> with batch size <em>n*b</em>.</li> |
| <li><code>dist_device_sync</code> is similar to <code>dist_sync</code>. The difference between them is that |
| <code>dist_device_sync</code> aggregates gradients and updates weight on GPUs |
| while <code>dist_sync</code> does so on CPU memory.</li> |
| <li><code>dist_async</code> performs asynchronous updates. |
| The weight is updated whenever gradients are received from any machine. |
| The update is atomic, i.e., no two updates happen on the same weight at the same time. |
| However, the order is not guaranteed.</li> |
| </ul> |
| |
| <h3 id="how-to-launch-a-job">How to Launch a Job</h3> |
| |
| <blockquote> |
| <p>To use distributed training, we need to compile with <code>USE_DIST_KVSTORE=1</code> |
| (see <a href="/get_started">MXNet installation guide</a> for more options).</p> |
| </blockquote> |
| |
| <p>Launching a distributed job is a bit different from running on a single |
| machine. MXNet provides |
| <a href="https://github.com/apache/mxnet/blob/v1.x/tools/launch.py">tools/launch.py</a> to |
| start a job by using <code>ssh</code>, <code>mpi</code>, <code>sge</code>, or <code>yarn</code>.</p> |
| |
| <p>An easy way to set up a cluster of EC2 instances for distributed deep learning |
| is using an <a href="https://github.com/awslabs/deeplearning-cfn">AWS CloudFormation template</a>. |
| If you do not have a cluster, you can check the repository before you continue.</p> |
| |
| <p>Assume we are at the directory <code>mxnet/example/image-classification</code> |
| and want to train LeNet to classify MNIST images, as demonstrated here: |
| <a href="https://github.com/apache/mxnet/blob/v1.x/example/image-classification/train_mnist.py">train_mnist.py</a>.</p> |
| |
| <p>On a single machine, we can run:</p> |
| <div class="highlight"><pre><code class="language-bash" data-lang="bash">python train_mnist.py <span class="nt">--network</span> lenet |
| </code></pre></div> |
| <p>Now, say we are given two ssh-able machines and <em>MXNet</em> is installed on both machines. |
| We want to train LeNet on these two machines. |
| First, we save the IPs (or hostname) of these two machines in file <code>hosts</code>, e.g.</p> |
| <div class="highlight"><pre><code class="language-bash" data-lang="bash"><span class="nv">$ </span><span class="nb">cat </span>hosts |
| 172.30.0.172 |
| 172.30.0.171 |
| </code></pre></div> |
| <p>Next, if the mxnet folder is accessible from both machines, e.g. on a |
| <a href="https://help.ubuntu.com/lts/serverguide/network-file-system.html">network filesystem</a>, |
| then we can run:</p> |
| <div class="highlight"><pre><code class="language-bash" data-lang="bash">python ../../tools/launch.py <span class="nt">-n</span> 2 <span class="nt">--launcher</span> ssh <span class="nt">-H</span> hosts python train_mnist.py <span class="nt">--network</span> lenet <span class="nt">--kv-store</span> dist_sync |
| </code></pre></div> |
| <p>Note that here we</p> |
| |
| <ul> |
| <li>use <code>launch.py</code> to submit the job.</li> |
| <li>provide launcher, <code>ssh</code> if all machines are ssh-able, <code>mpi</code> if <code>mpirun</code> is |
| available, <code>sge</code> for Sun Grid Engine, and <code>yarn</code> for Apache Yarn.</li> |
| <li><code>-n</code> number of worker nodes to run on</li> |
| <li><code>-H</code> the host file which is required by <code>ssh</code> and <code>mpi</code></li> |
| <li><code>--kv-store</code> use either <code>dist_sync</code> or <code>dist_async</code></li> |
| </ul> |
| |
| <h3 id="synchronize-directory">Synchronize Directory</h3> |
| |
| <p>Now consider if the mxnet folder is not accessible. |
| We can first copy the <code>MXNet</code> library to this folder by |
| <code>bash |
| cp -r ../../python/mxnet . |
| cp -r ../../lib/libmxnet.so mxnet |
| </code></p> |
| |
| <p>then ask <code>launch.py</code> to synchronize the current directory to all machines' |
| <code>/tmp/mxnet</code> directory with <code>--sync-dst-dir</code></p> |
| <div class="highlight"><pre><code class="language-bash" data-lang="bash">python ../../tools/launch.py <span class="nt">-n</span> 2 <span class="nt">-H</span> hosts <span class="nt">--sync-dst-dir</span> /tmp/mxnet <span class="se">\</span> |
| python train_mnist.py <span class="nt">--network</span> lenet <span class="nt">--kv-store</span> dist_sync |
| </code></pre></div> |
| <h3 id="use-a-particular-network-interface">Use a Particular Network Interface</h3> |
| |
| <p><em>MXNet</em> often chooses the first available network interface. |
| But for machines that have multiple interfaces, |
| we can specify which network interface to use for data |
| communication by the environment variable <code>DMLC_INTERFACE</code>. |
| For example, to use the interface <code>eth0</code>, we can</p> |
| <div class="highlight"><pre><code class="language-" data-lang="">export DMLC_INTERFACE=eth0; python ../../tools/launch.py ... |
| </code></pre></div> |
| <h3 id="debug-connection">Debug Connection</h3> |
| |
| <p>Set<code>PS_VERBOSE=1</code> to see the debug logging, e.g |
| <code> |
| export PS_VERBOSE=1; python ../../tools/launch.py ... |
| </code></p> |
| |
| <h3 id="more">More</h3> |
| |
| <ul> |
| <li>See more launch options by <code>python ../../tools/launch.py -h</code></li> |
| <li>See more options of <a href="http://ps-lite.readthedocs.org/en/latest/how_to.html">ps-lite</a></li> |
| </ul> |
| |
| </div> |
| </div> |
| |
| </div> |
| </div> |
| |
| </article> |
| |
| </main><footer class="site-footer h-card"> |
| <div class="wrapper"> |
| <div class="row"> |
| <div class="col-4"> |
| <h4 class="footer-category-title">Resources</h4> |
| <ul class="contact-list"> |
| <li><a href="/versions/1.9.1/community/contribute#mxnet-dev-communications">Mailing lists</a></li> |
| <li><a href="https://cwiki.apache.org/confluence/display/MXNET/Apache+MXNet+Home">Developer Wiki</a></li> |
| <li><a href="https://issues.apache.org/jira/projects/MXNET/issues">Jira Tracker</a></li> |
| <li><a href="https://github.com/apache/mxnet/labels/Roadmap">Github Roadmap</a></li> |
| <li><a href="https://medium.com/apache-mxnet">Blog</a></li> |
| <li><a href="https://discuss.mxnet.io">Forum</a></li> |
| <li><a href="/versions/1.9.1/community/contribute">Contribute</a></li> |
| </ul> |
| </div> |
| |
| <div class="col-4"><ul class="social-media-list"><li><a href="https://github.com/apache/mxnet"><svg class="svg-icon"><use xlink:href="/versions/1.9.1/assets/minima-social-icons.svg#github"></use></svg> <span class="username">apache/mxnet</span></a></li><li><a href="https://www.twitter.com/apachemxnet"><svg class="svg-icon"><use xlink:href="/versions/1.9.1/assets/minima-social-icons.svg#twitter"></use></svg> <span class="username">apachemxnet</span></a></li><li><a href="https://youtube.com/apachemxnet"><svg class="svg-icon"><use xlink:href="/versions/1.9.1/assets/minima-social-icons.svg#youtube"></use></svg> <span class="username">apachemxnet</span></a></li></ul> |
| </div> |
| |
| <div class="col-4 footer-text"> |
| <p>A flexible and efficient library for deep learning.</p> |
| </div> |
| </div> |
| </div> |
| </footer> |
| <footer class="site-footer2"> |
| <div class="wrapper"> |
| <div class="row"> |
| <div class="col-3"> |
| <img src="/versions/1.9.1/assets/img/asf_logo.svg" class="footer-logo col-2"> |
| </div> |
| <div class="footer-bottom-warning col-9"> |
| </p><p>"Copyright © 2017-2022, The Apache Software Foundation. Licensed under the Apache License, Version 2.0. Apache MXNet, MXNet, Apache, the Apache |
| feather, and the Apache MXNet project logo are either registered trademarks or trademarks of the |
| Apache Software Foundation."</p> |
| </div> |
| </div> |
| </div> |
| </footer> |
| |
| |
| |
| |
| </body> |
| |
| </html> |
